Automated Model Management Methods

1. A method for iterative model deployment, comprising: receiving raw data from a set of data sources; extracting features from the raw data; labeling the raw data; at a deployed event detection model within an event detection system, detecting an occurrence of an event class based on the features; in response to detecting the occurrence of the event class, transmitting a notification to an endpoint associated with the event class; executing a model generation method at a testing platform, comprising: identifying raw data with labels pre-associated with the event class; segmenting the identified raw data into a training data set and a testing data set, wherein the training data set is disjoint from the testing data set; generating a candidate model for the event class by training a seed model for the event class with the training data set; determining evaluation metric values for the candidate model using the testing data set; and in response to the evaluation metric values satisfying a set of deployment conditions, deploying the candidate model to the event detection system, comprising detecting a subsequent occurrence of the event class based on subsequent raw data received from the set of data sources, using the candidate model.

2. The method of claim 1 , further comprising detecting concept drift from the raw data, wherein the model generation method is automatically executed in response to detecting the concept drift.

3. The method of claim 1 , wherein the model generation method is executed at an execution frequency, wherein the execution frequency is dependent on a concept drift rate associated with the event class.

4. The method of claim 3 , wherein the concept drift rate is determined from a rate of change in an inaccuracy rate of the deployed event detection model.

5. The method of claim 4 , wherein the inaccuracy rate of the deployed event detection model is determined based on whether an expected subsequent event occurs within a predetermined time interval after detecting a precursor event associated with the event class.

6. The method of claim 3 , wherein the concept drift rate is determined for a geographic region, wherein the candidate model is specific to the geographic region.

7. The method of claim 1 , wherein the raw data is labeled by a labeling model, wherein the method further comprises generating the labeling model, comprising: extracting secondary features from secondary training data, wherein the secondary training data is pre-associated with the label; and training a seed labeling model to algorithmically arrive at the label.

8. The method of claim 1 , wherein the occurrence of the event class is detected when a probability of a combination of feature values for the features, extracted from raw data associated with a geographic region, falls below a threshold probability.

9. The method of claim 1 , wherein the identified raw data comprises raw data from multiple modalities.

10. The method of claim 1 , further comprising: tracking a number of notifications, generated by the deployed event detection model, that are used by the endpoint; calculating a preliminary metric based on the number; transmitting subsequent notifications to the endpoint in response to detecting subsequent occurrences of the event class with the candidate model; tracking a second number of the subsequent notifications used by the endpoint; and calculating a candidate conversion metric based on the second number; wherein the deployed event detection model is replaced with the candidate model when the candidate conversion metric exceeds the preliminary metric.

11. The method of claim 1 , wherein the candidate model is associated with a geographic region, wherein the subsequent raw data comprises data geotagged with geographic locations within the geographic region.

12. The method of claim 11 , further comprising: deploying a second candidate model for the event class concurrently with the candidate model, the second candidate model associated with a second geographic region different from the geographic region; detecting a second subsequent occurrence of the event class, using the second candidate model, based on secondary subsequent raw data associated with secondary geographic locations within the second geographic region.

13. The method of claim 1 , further comprising deploying a second candidate model for a second event class concurrently with the candidate model.

14. The method of claim 1 , wherein the occurrence of the event class is detected in near-real time.

15. A model management method, comprising: detecting concept drift associated with an event class; based on the concept drift, automatically iterating through: determining a data pool comprising data labeled with labels pre-associated with the event class; segmenting the data pool into a training data set and testing data set; training a seed model with the training data set to generate an event detection model for the event class; determining evaluation metrics for the event detection model using the testing data set; deploying the event detection model within a production environment in response to the evaluation metrics satisfying deployment conditions; and detecting an occurrence of the event class based on raw data received from a set of data sources using the event detection model.

16. The method of claim 15 , wherein detecting concept drift comprises detecting concept drift from the raw data.

17. The method of claim 15 , wherein determining a data pool comprises determining a data pool that includes one or more of: image data, text data, video data, or audio data.

18. The method of claim 15 , wherein deploying the event detection model within a production environment comprises replacing a previously deployed model with the event detection model within the production environment.

19. A model management method, comprising: automatically iterating through: determining a data pool comprising data labeled with labels pre-associated with an event class; segmenting the data pool into a training data set and testing data set; training a seed model with the training data set to generate a candidate model for the event class; determining evaluation metrics for the candidate model using the testing data set; deploying the candidate model within a production environment in response to the evaluation metrics satisfying deployment conditions; detecting an occurrence of the event class based on raw data received from a set of data sources using the candidate model; identifying subsequent raw data geotagged with geographic locations within a geographic region; determining feature values from the subsequent raw data; and detecting an occurrence of an event based on the feature values.

20. The method of claim 19 , wherein detecting an occurrence of an event comprises detecting one or more of: an event start time, an event end time, or an event location.